In this paper we introduce a novel supervised algorithm for equine activity recognition based on accelerometer data. By combining an approach of calculating a wide variety of time-series features with a supervised feature significance test we can obtain the best suited features using just 5 labeled samples per class and without requiring any expert domain knowledge. By using a simple cluster assignment algorithm with these obtained features, we get a classification algorithm that achieves a mean accuracy of 90+%. In this paper we will compare this approach to a state-of-the-art convolution neural network classifier both in terms of accuracy as well as in terms of number of labeled samples that were used to train the classifier.